基于两阶段的轻量级深度估计方法

打开文本图片集

中图分类号：TP183 文献标志码：A

Abstract： To address the issues in traditional monocular depth estimation -where relative estimation methods lose scale information， metric estimation methods suffer from insufficient edge precision， and existing depth networks have large parameter counts and high computational costs-a two-stage fusion framework called LacDepth was proposed. This framework aimed to fuse metric and relative estimation methods through two stages. In the first stage，the deep residual pyramid module adopted a multi-scale Laplacian residual compensation mechanism and effectively improved the geometric fidelity of edge contours via a high-frequency feature enhancement strategy. In the second stage，the lightweight atractor-driven classifier constructed a three-level cascaded depth interval prediction network， established a pixel-level probability density function based on the conditional logbinomial distribution， and realized sub-interval fine-tuning of relative depth values through differentiable weighting. Experimental results show that LacDepth achieves the best comprehensive performance on the KITTI dataset， with an average relative error of 0.059 and a parameter count of 9.8×106 ， demonstrating significant advantages in balancing precision and efficiency.

Keywords： depth estimation; multi-scale feature fusion; feature enhancement; lightweight network; classification strategy

在计算机视觉领域，单图像深度估计（singleimagedepthestimation，SIDE）一直是一个备受关注的问题。（剩余18548字）